Suspicions arise over "AI Central Bank" or "Balance Sheet as a Service"? Nvidia halts "Revenue Sharing for Financing" model

Nvidia is paying the price for its ambitious AI ecosystem financing plans. Less than two months after announcing its “AI Compute Partnership Program,” the world’s highest-valued company quietly suspended some deals, just as its off-balance-sheet commitments and guarantees reached a staggering $530.5 billion—an amount revealed for the first time and raising nerves among credit market investors.

According to The Wall Street Journal on August 27th, sources revealed that Nvidia last week suspended certain agreements under the revenue sharing plan, partly because employees were concerned that the model might trigger antitrust scrutiny and create uncertainty over the boundaries of the chip giant’s operational control over customers.

Meanwhile, Morgan Stanley initiated credit coverage on Nvidia for the first time, assigning a neutral rating and warning that the company is transforming its balance sheet into a strategic AI financing tool—so-called “Balance-Sheet-as-a-Service”—which introduces new tail risks.

Analysts believe these two developments together have sharply intensified Wall Street’s skepticism of Nvidia’s “circular financing” model. Nvidia’s credit default swap (CDS) quotes are near historical highs, closing at 83 basis points, more than double the historical average.

Morgan Stanley credit analyst Lindsay Taylor stated bluntly that Nvidia’s over $500 billion in partnerships are still at the memorandum of understanding stage, with revenue sharing and credit support frameworks barely disclosed, and the tail risk is “too early, too opaque, and too large in scale.”

Behind the Halt: Antitrust Concerns and Power Struggles

According to The Wall Street Journal, Nvidia announced its “AI Compute Partnership” in July this year, with a core logic to solve financing difficulties for small and medium-sized cloud service providers.

Building AI cloud infrastructure requires upfront investment of billions to buy Nvidia GPUs and data centers, while these service providers often need financing before signing enough customer contracts. Nvidia’s solution: promise to rent their GPU capacity itself if the service provider can’t find other customers, ensuring a guaranteed income source that helps them secure loans for infrastructural development.

The plan also opens two profit streams for Nvidia: first by selling chips, and second by taking a share of income from clients renting out those chips. According to sources, the basic hourly rate is set to cover service providers’ costs (including chip depreciation, data center operations, and staff), and Nvidia will collect 50% of the income exceeding that threshold. The first cloud service providers to join were Sharon AI and Firmus Technologies.

However, the scheme ran into friction within weeks of its launch. Sources say Nvidia demanded that service providers only rent chips to approved customers, and preferred distributing capacity to multiple small AI companies instead of leasing to a single large customer. Many service providers strongly resisted, believing client selection is a basic commercial right. Nvidia staff subsequently warned existing and potential customers of antitrust concerns, noting the sensitivity of the chip giant’s extent of intervention in clients’ decisions.

Nvidia formally suspended some deals under the plan last week, less than two months after its launch. A Nvidia spokesperson said, “The new business model launched in July remains effective and is continuously evolving due to strong demand,” without denying the suspension. Sources say Nvidia may revise or merge the program into other projects in the future.

Off-Balance-Sheet Commitments Exposed: The $530.5 Billion Shocks the Market

In the same week as the suspension news, Nvidia disclosed the specific scale of its off-balance-sheet commitments in its latest quarterly financial report, drawing market attention with the sheer numbers.

According to estimates by Australian Financial Review (AFR), Nvidia’s obligations cover several categories:

Commitments to purchase storage capacity from suppliers surged from $119 billion to $279 billion within three months; commitments for providing land, power, and data center support capacity to new service providers like Firmus and Iren reached $56 billion; loan guarantees to other datacenter developers amounted to $108.5 billion.

In total, Nvidia disclosed $530.5 billion in off-balance-sheet commitments and guarantees, with durations extending beyond 2032.

Nvidia CFO Colette Kress responded during the earnings call to external questions about circular financing:

“We’re aware of the scale of these commitments, and we know some will call this circular financing. But we see it differently. We are experiencing a major computing platform transition—one of the most important technological shifts in human history—with once-in-a-lifetime companies.”

She also predicted that this new income stream could generate tens of billions in profits for Nvidia over the medium to long term.

However, financial data also revealed another risk: Q2 net profit was $56 billion, while operating cash flow was only $24 billion, leaving a $36 billion gap.

This gap mainly stems from a sharp increase in accounts receivable—by the end of the reporting period, Nvidia’s accounts receivable balance reached $63.1 billion, up sharply from $38.5 billion in January. Nvidia admitted in the filing, “Financing arrangements with certain investment-grade customers, including extended payment terms under large, multi-quarter agreements, will continue to impact the timing of operating cash flows.”

Significantly, just five companies (believed to be Microsoft, Amazon, Alphabet, Meta, and Oracle) accounted for 70% of those accounts receivable, highlighting high concentration risk.

Morgan Stanley Dissects “Tail Risks”: Three Major Exposures Deconstructed

In its first Nvidia credit coverage report, Morgan Stanley built a framework going beyond traditional leverage metrics, deconstructing Nvidia’s credit exposures into three main classes and forecasting that by end-2028, Nvidia’s total credit exposures may reach $200 billion, including around $170 billion from various adjustments and contingent obligations.

First Class: Lease Liabilities and Guarantee Exposures

Nvidia disclosed $32.4 billion in lease obligations yet to begin, with tenors of 3 to 20 years, expected to commence between fiscal 2027 Q2 and fiscal 2033. Morgan Stanley estimates that these commitments may result in an extra $20.8 billion in lease liabilities.

Additionally, Nvidia disclosed $3.5 billion in maximum exposure to partner facility lease guarantees, and provided a residual value guarantee (RVG) covering about 4.25 gigawatts of IT load for the PORTS-Pike Technology Park in Ohio.

S&P expects that as related facilities come online between 2028 and 2030, Nvidia’s debt adjustments will rise from $4.2 billion in 2028 to about $37.7 billion in 2031–2032.

Second Class: Residual Value Support under the $500 Billion Partnership Plan

In August, Nvidia announced collaborations with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, aiming to mobilize over $500 billion in third-party capital for AI infrastructure. Nvidia said in some projects it may provide residual value support up to 25%.

Morgan Stanley developed a scenario-driven, multi-period model, rolling this framework out to about 15 financing batches of $5 billion each, estimating Nvidia’s maximum contingent liability at about $90 billion, peaking shortly after the end of 2028.

The firm noted that the $500 billion partnership plan is still at a memorandum of understanding stage, with high uncertainty around financing size, pace, and guarantee structures; as exposures get transferred via special purpose vehicles (SPVs) and private placements, disclosure may become quite limited in the future.

Third Class: Revenue Sharing and Unsold Compute Shortfall Guarantees. Morgan Stanley deems this category potentially most far-reaching and hardest to track. Its scenario model of 5 gigawatts of supported compute estimates contingent liabilities of about $81 billion (about $64 billion after tax) by year-end 2028.

The most direct precedent is Nvidia’s $6.3 billion commitment to CoreWeave—to buy its excess cloud compute if CoreWeave can’t find other customers. S&P has included such unsold compute guarantees in debt adjustments, treating them like lease guarantees.

Credit Market Disagreements: Fundamentals Robust, but Tail Risks Underpriced

Morgan Stanley’s analysis presents inherent tension: fundamental data remain robust, but the opacity of tail risk prevents a buy recommendation.

On fundamentals, even after fully including all $170 billion adjustments and contingent obligations, Nvidia’s balance sheet remains extremely strong. Morgan Stanley predicts Nvidia’s net cash (post balance sheet debt) can reach around $520 billion by fiscal 2029/2028 calendar year, or about $350 billion after full adjustment. All-in leverage is just 0.4x, with post-shareholder distributions, free cash flow in excess of total all-in debt. Even if EBITDA and free cash flow stall at 2027 forecast levels, these measures would still stand at about 0.7x and 15% respectively by 2029.

However, credit market pricing has subtly changed. Nvidia’s 5-year CDS is around 84–89 basis points, roughly 25 basis points wider than equivalently rated Alphabet and Amazon, while the spread on Nvidia’s 2036 bond is only 10–20 basis points tighter than what lower-rated IBM and AT&T bonds pay.

Nick Ferres, CIO of Vantage Point Asset Management, points out, “How you interpret Nvidia’s performance depends on whether you’re an equity or a credit investor—for the latter, the extra layers of risk from circular financing always loom large.”

Morgan Stanley advises investors to remain “on the sidelines” for both cash bonds and CDS, adding it would consider getting involved only if CDS breaches 100 basis points or cash bond spreads approach AT&T levels. The bank also cautions that CDS may underperform cash bonds due to lender hedging and thematic positioning.

Nvidia’s Defense: “This Time is Really Different”?

Faced with comparisons to the fiber-optic giants’ “lending money to buyers of their own products” during the dotcom bubble, Nvidia management chose to confront the doubts head-on.

Colette Kress cited projected capex for hyperscale cloud providers during the earnings call: the five largest operators are expected to invest nearly $800 billion in 2026 and $1.3 trillion in 2027.

CEO Jensen Huang defended sustained growth from the demand side, stressing that AI is shifting from human- to AI agent-led, “the compute demand for an agent is 15 to 100 times that of a human—the scale of required compute is extraordinary.”

He also made a rare move by providing preliminary full-year guidance for fiscal 2028, projecting revenue growth of 70%—far above the previously expected consensus of 45%—which directly triggered a share price rebound after five consecutive quarters of “beats but falls.”

However, financial blog ZeroHedge warns that “this time is different” remains the most dangerous phrase in financial history. Whether Nvidia’s circular financing logic holds depends ultimately on whether AI demand fulfills management’s projections. Morgan Stanley’s conclusion perhaps best captures the current mainstream mood of the credit market:

Nvidia’s fundamentals are impeccable, but before over $1 trillion in GPU and XPU-related ecosystem financing is rolled out via innovative structures, whether this balance sheet is a moat or a hidden risk remains a question only time can answer.